LASER SCANNING OR IMAGE-BASED MODELING? A COMPARATIVE THROUGH THE MODELIZATION OF SAN NICOLAS CHURCH

D.G. Aguilera, J. G. Lahoz

Land and Cartography Engineering Department, Univ. of Salamanca, Spain [email protected], [email protected]

KEY WORDS: Cultural Heritage, Laser scanning, Photogrammetry, Image-based modeling

ABSTRACT:

Nowadays, the demand for 3D models of historical monuments is continuously growing in the field of archaeological and architectural applications. Currently, the two main sources that can provide detailed and reliable 3D surface models are Photogrammetry through image-based modeling and Terrestrial Laser Scanner through laser scanning techniques. Among the plenty of works so far presented the use of laser scanner for Cultural Heritage survey seems to aim towards a monopoly in the 3D modeling pipeline. Nevertheless, up to now some geometric related issues have been not yet solved by laser scanner and thus, time and patience are required to get final results. Consequently, Photogrammetry through image-based modeling seems to be an economical and efficient alternative to accomplish these problems, especially when close range applications involving simple objects and some geometrical constraints can be assumed.

According to such remark, in this paper we deal with the comparison between image-based modeling and laser scanning techniques applied to the survey of the outside of the emblematic romanic church of San Nicolas, located in the city of Avila (Spain). Results and conclusions have been reported in the end to answer the question: laser scanning or image-based modeling?. To this aim, the work was carried out according to three different stages. Firstly, the church was fully surveyed with a terrestrial laser scanner and post-processed using laser scanning software. In a second stage, digital colour images were acquired with a high-resolution camera and post-processed using image-based modeling software. Finally, in a third stage a global comparison of both 3D models based on results, accuracy and time was assessed.

1. INTRODUCTION As a result, the question, laser scanning or image-based modeling?, can not be answered across the board. As we have In the Cultural Heritage avenue, the classical and mature seen, each technique owns its advantages and disadvantages at Photogrammetry has experimented an important evolution: different working fields and depending on the own object’s from 3D coordinate extraction of points through stereo- features. Even in many cases a combination of both techniques restitution to Digital Photogrammetry based on image-based might be a useful solution. Approaches such as (Finat et. al. modeling which exploits the bundle adjustment technique and 2005) and (El-Hakim et. al., 2005) have been addressing the use the overlapping between images. Nevertheless, the number of of range scans and images to develop the modeling of complex produced points and the level of automation for close range sites. applications involving complex objects are still relatively low and require time consuming procedures. In this sense, our goals overlap with the work remarked of these researchers, but before that we want to perform a comparative Although the development of specialized software for direct focused on qualitative and quantitative analysis of these two production of 3D models following simple photogrammetric methodologies. techniques have broadened the field of applications, yet this does not change the basic characteristics of the method. The Next sections deal with the approaches developed in San emergence of terrestrial laser scanning and state-of the art Nicolas Church for 3D modeling from laser scanning and software developments for processing large amount of the image-based modeling. A final section is devoted to show a produced data may lead to the impression that this technology is comparative between both approaches extracting conclusions the main solution for 3D models generation. However, laser and results. scanners still remain quite bulky instruments and difficult to use especially for data collection entailing high level positions with respect to the ground. Consequently, the existence of occlusions 2. HISTORICAL AND ARCHITECTURAL ASPECTS and the lack of data in some parts of the objects are frequent, ABOUT SAN NICOLAS CHURCH resulting to incomplete models. Moreover, although there is certain autonomy of laser scanning with relation to topographic The origins of the church of San Nicolás (Figure 1) date back to field work, still this is not always the case; i.e., when there is a the early days of the repopulation of Castille, a fact supported need to incorporate a model into a particular reference system, by an inscription which shows that its devotion to San Nicolás, or when the shape and characteristics of the monument do not the Bishop, took place in 1198. It is right in the middle of the allow a reliable cloud-to-cloud registration, there is need of Moorish quarter, 200 m outside the City Walls, and the poverty surveying measurements. of its materials reflects in some way the poverty of its former parishioners.

modelling of closed surfaces are receiving great attention as they are not completely solved. Moreover, the challenge for surface reconstruction algorithms lies in finding methods which can cover a wide variety of shapes.

The scanning process at each station consisted on the acquisition of a low resolution RGB image with the scanner camera in order to manage the areas to be scanned and also to project the image texture on the point cloud. A medium-range laser scanner was used. Trimble GS200 (Figure 2), which incorporates a rotating head and two inner high speed rotating mirrors that allow to acquire a scene with a large enough field of view, i.e. 360º H x 60º V, reducing the need of using lots of scan stations. The sensor accuracy is below 1.5mm at 50m of distance with a beam diameter of 3mm. In our case, a resolution of 30 mm was set up in both directions (horizontal and vertical) at 30 meters of distance. Figure 1: San Nicolas’ Church

From the outside it can be seen that the church has a single apse, a tower and two doors. The tower has three decreasing sections, and was built in two different periods, which explains why the access staircase is spiral in the first section, and made of wood supported by the walls in the upper half. Apparently, it never had bells.

The church is 30 m long (east-west direction), 22 m wide and 12 m tall in the central part with a maximum height at the tower with 24 m. This fact leads to an extra laser station, away enough of church in order to scan the upper part of the tower due to the limited scanning vertical range of Trimble GS200 (40º above the horizon).

It is formed by a three naves main body, articulated with the apse, the tower and a minor construction devoted to the sacristy. It shows the typical romanic style absence of ornamentation except for the two doorways. They are placed at the north and south facades with five and two archivolts respectively and the also usual decorative motives that demand a higher resolution Figure 2. Terrestrial laser scanner: Trimble GS200 scan. The main nave exhibits four buttresses in its occidental (www.trimble.com). part. After that, the data post-processing step included: The church is in the middle of a city garden and so, it makes easier to surround it and place the laser scanner or whatever • Cloud point depuration applying spatial and sensor. The church is relatively well preserved, i.e. it shows a topographic filters to eliminate overlapping areas and high regularity in its surfaces geometry and radiometry. to erase noisy information. Manual segmentations were also applied to eliminate those features that were On the side of the drawbacks, we may talk of a dense “curtain” not eliminated by the filters. of trees which occlude part of the facades. Both on the east and north parts, buildings are too close to the church to allow a • A semiautomatic register between point’s clouds flexible design of the geodetic network. On the west there is a based on ICP (Iterative Closest Point) algorithm and high traffic street while at the east there is a quiet street but full supported by the search of point’s correspondences. of middle size vehicles parked there. • Mapping of high resolution textures. An external digital camera is registered (manually) to laser model 3. 3D MODELING FROM LASER SCANNING in order to obtain high resolution RGB values projected onto the points cloud. Surface reconstruction is a well studied problem in Computer Graphics with a wide range of applications. With the advent of • Derivate products: mesh generation, sections, laser scanner systems, which can provide dense data sets from a orthophotos, fly through and simplified models. variety of objects, the issues of surface reconstruction and

Nowadays, all above operations are performed in a number of For the comparison described in this paper, a high-resolution modeling software packages (e.g. RapidForm, Cyclone, camera, Nikon D70 (Figure 4), was used to capture digital Polyworks, RealWorks, etc). In our case, RealWorks Survey 5.1 images, while in the 3D modelling approach PhotoModeler was used to accomplish these tasks. Nevertheless, the main software package (www.photomodeler.com) was used. problem associated to these tools is that none allow to obtain efficient simplified models which incorporate topological PhotoModeler is a powerful 3D software product that calculates relations together with geometrical constraints, providing an measurements and constructs 3D models from images simply interactive and virtual representation on Internet. Therefore, and easily. This software has the ability of: there exist a big gap between laser scanner software and powerful modelling rendering and visualization packages • Self-calibration or introduction of interior orientation (known as ‘computer animation software’), which use polygons parameters, allowing the use of non-metric cameras. to accurately represent the model, providing an optimal surface • The use of lines between points for the determination description. of the delineation of the objects. • Imposing constraints, such as collinearity or coplanarity of points, perpendicular or parallel lines. 4. 3D MODELING FROM IMAGES • Determining epipolar lines improving the location of homologue points. In the last years, the development of systems and specific • Producing models without control points, while the software tools in the field of Close Range Photogrammetry have scale is determined by measured distances. experimented a huge proliferation. These software packages • Adding to the model mathematical surfaces, such as incorporate photogrammetric algorithms which allow to solve basic primitives (cylinders, cones, etc.) or other image-based modeling approaches without stereoscopic vision, second order surfaces (quadrics). using bundle adjustment together with geometrical constraints. • Producing orthoimages at defined projection planes. By pointing manually and monoscopically on homologue points • Creating TIN and wireframe models. on more than two overlapping images the optimum ray • Appying texture to the model from images selected intersections are determined. For better accuracy and more manually or automatically, and producing of reliable data determination, images should be taken rather with photorealistic textured models. a convergence of 20º-90º (Figure 3), instead of the desired • 3D viewer, with zoomimg, rotating and measuring normal case of the convetional procedure of stereo-restitution. capacity on the model. The creation of the model is achieved by selecting points that create planes or other mathematical surfaces (basic primitives) and by adding geometric constraints. A big advantage of these packages, e.g. PhotoModeler, 3D Builder, Canoma and others, is the simplified and user friendly interface that they offer, since they are made mainly for non-photogrammetrist. The basic problems are the lack of automated procedures, for example matching, which increases the required workload, and the insufficient ways for the assessment of the achieved accuracy. The procedures and technical information i.e. mathematical model used or radial lens distortion approach, remain hide like a ‘black box’. However, it is a cost effective solution with interactive processing for the 3D geometric recording of Figure 4. Digital camera: Nikon D70 monuents and without any need for special knowledge of (www.nikon.com). Photogrammetry. Through these capabilities, local or full 3D models using one or more digital cameras or even multiple frames of video recording of the monument can be generated. Speed, ease, low- cost, and the variety of alternatives in data acquisition make the use of such types of software very attractive. However, the production of detailed and accurate models of a complex building or monument can be a cumbersome and time consuming task. The topography of the surrounding area of the monument and the existence of possible obstacles may hinder the multiple image capturing and thus the effectiveness of the procedure. However, the automatic creation of 3D compact models with topology and geometrical constraints and without holes or gaps is an adventage.

Figure 3. Overlapping images taken in the convergence case.

5. CASE STUDY: SAN NICOLAS CHURCH Table 1. Statistic data related with mathematical model in image-based modelling approach. In this paragraph we report a qualitative and quantitative analysis between image-based modeling and laser scanning With relation to the most important step, the processing, an techniques applied to the survey of the outside of the image-based modelling approach using PhotoModeler was emblematic romanic church of San Nicolas. performed. This step includes the 3D modelization with textures, as well as the restitution and vectorization of each Following the 3D modeling from images approach several architectural element (Figure 6). A workload of 150 hours was aspects related to accuracy and time are remarked: required to accomplish this task. An average accuracy of 3.4 cm was obtained through this methodology. In the camera self-calibration step, a rectangular and two- dimensional grid was used as a testing bench formed by four control points and 107 tie points (Figure 5).

Figure 6. Wireframe model of San Nicolas church obtained through semi-automatic restitution and vectorization.

Figura 5. Calibration grid used in camera self-calibration. Finally, with the goal of promoting and publish our work a interactive environment based on VRML (Virtual Reality This step carried out a workload of 4 hours to photographic Modeling Language) was developed, in which the user can fell taking and 2 hours to calibration processing. The internal immersed and even touch the different elements of the church parameters obtained were compared with other calibration (Figure 7). Furthermore, alphanumeric information was packages (such as Pictran and Matlab), obtaining good results attached to the model giving as a result a Geographical in focal length and principal point coordinates and small Information System applied to the Cultural Heritage. This task discrepancies in radial lens distortion coefficients. consumed a total of 150 hours.

After that, a photographic taking of San Nicolas church was performed consuming a total of 4 hours. A total of 106 images were acquired from different viewpoints around the church with the aim of guarantying a perfect coverage of the object, obtaining enough information in object modelling. A mathematical model composed by 157.636 equations and 78.636 unknowns was conformed (Table 1), guarantying enough number of redundancies, 79.000.

IMAGES EQUATIONS UNKNOWNS Fixed: Internal camera parameters

Variables: 106 6x106img=636 Figure 7. Textured model of San Nicolas church, adding a IMAGES eqs. (camera virtual recreation of the surroundings. viewpoints) On the other hand, following the 3D modeling from laser Tie Points scanner approach the next requirements were taken into Redundancies: r: 79.000 26.000ptos.x3eq= account: 78.000 unknowns Total: 78.636 Total: 157.636 equations unknowns.

In the acquisition process, a minimal number of 8 stations were Derivate products such as: mesh generation, orthophotos or used to guarantee a good coverage of the objects and reduce metric measurements were immediate, requiring 4 hours to get error propagation between scans. A resolution of 30 mm x 30 them. However, maps generation i.e. wireframe models or mm for a distance of 20 meters was set up. isometric views required time and patience for obtaining a final result. A total of 150 hours were used to obtain complete The total time used in the scanning process was about 6 hours isometric views. with an average time of 45 minutes per station, obtaining 5 millions of points scanned. A relative accuracy of 2 mm was Next table (Table 2) shows a comparison between both obtained in each point scanned. approaches based on five aspects: accuracy, time, cost, results and applicability. After that, a cloud point treatment was carried out using spatial and topographic filters to reduce information and to eliminate Accuracy Time Cost Results Applicability noisy and unnecessary data (Figure 8). Manual segmentations Image-based 300 Very 3.4 cm 310h Total were also applied to eliminate those features that were not modelling €/day good Laser 1.500 eliminated by filters. This task consumed 2 hours. 1.3 cm 172h Good Medium Scanning €/day The register of multiple clouds based on the search of homologous points carried 2 hours. After alignment process a Table 2. Quantitative and qualitative aspects about image-based global accuracy of 1.3 cm was obtained. modeling and laser scanning.

6. DISCUSION

In this paper we have dealt with the comparison between two different approaches in 3D modeling of Cultural Heritage: the traditional image-based modeling and the new technology of laser scanning. We have reached clear conclusions about the effectiveness of both techniques for Cultural Heritage modeling.

From a quantitative point of view, laser scanning approach provided more accurate results than image-based modelling. Furthermore, time required in laser scanning was only of 172 hours, while image-based modelling needed 310 hours. Nevertheless, several qualitative aspects have to take into Figure 8. Alignment of different scans and depuration of account in both methodologies. With relation to results, image- unnecessary information. based modelling offered great results with a wide range of applications: from an interactive environment to metric maps Finally, a manual procedure based on mapping high resolution and wireframe models ready to be plotted. Efficient simplified textures was performed. An external digital camera was models which incorporate topological relations together with registered to laser model identifying correspondences between geometrical constraints were created, providing virtual 3D points cloud and 2D image points. As a result, a final laser representations on Internet through VRML. Furthermore, model with high resolution RGB values projected onto the image-based modeling constitutes a low-cost technique while points cloud was obtained (Figure 9). This task was especially laser scanning requires a large investment. delicate due to the different illumination conditions present in digital images. A total of 8 hours were required to map high- As a result, the question which technique is ‘better’ than the resolution textures properly. other can not be answered across the board. It will be a mix of factors and elements such as own characteristics of the object, applicability, economical support, client requirements, etc, which will provide the best strategy. As we can see (Table 3), each technique owns its advantages and disadvantages at different working fields. In many cases, a combination of both techniques might be a useful solution.

3D modeling from laser scanner 3D modeling from images ↓ No semantic information ↑ Semantic information ↓ Inaccurate lines and joints ↑ Accurate lines and joints ↓ Poor colour information ↑Good colour information ↑ Prompt and accurate metric ↓ Hard-working and slow information metric information ↑ Potency and automatization in ↓ Low potency and

Figure 9. Final laser model obtained mapping high resolution textures.

data capture automatization in data capture of International Workshop on Recreating the Past -Visualization ↑ Excellent technique for the ↓ Time-consuming technique and Animation of Cultural Heritage, Ayutthaya, Thailand. description of complex and for the description of complex irregular surfaces and irregular surfaces Ikeuchi K., Nakazawa A., Nishino K. and Oishi T. Creating ↓ High-cost technique ↑ Low-cost technique virtual Buddha statues through observation. In IEEE Workshop ↓ The 3D model is an entity ↑The 3D model is an entity on Applics. Of Comp. Vison in Architecture, 2003. disorganized and without topology organized and with topology ↑ Light is not required to work ↓ Light is required to work Karras, G., E., Mavrommati, D., Madani, M., Mavrelis, G., Lymperopoulos, E., Kambourakis, A., Gesafidis, S., 1999. Table 3. Comparison of features: Laser scanning vs. Image- Digital orthophotography in archaeology with low-altitude based modeling. non-metric images. ISPRS Vol XXII, Workshop ‘Photogrammetric measurement, object modeling and documentation in architecture and industry’, Thessaloniki, 7. REFERENCES Greece, 1999.

References from Journals:

Aguilera D. G., Gómez J., Finat J. A new method for vanishing ACKNOWLEDGEMENTS points detection in 3D reconstruction from a single view.

ISPRS Commission V, WG V/2. Mestre (Venice), 2005. First of all, we would like to express our gratitude to Soraya

Colino and Cristina Espadas for the performance of the image- Bernardini F., Rushmeier H., Martín I.M., Mittleman J., and based modeling work, providing good results and conclusions Taubin G. Building a digital model of Michelangelo’s about San Nicolas Church. Furthermore, we are especially Florentine Pieta. IEEE Comp. Graph. And Applic., 22(1):59- grateful with the University of Salamanca, especially with the 67, 2002. Extraordinary Courses Service for the partial economical

support through the development of Semminars and Summer Ioannidis, C., Potsiou, C., Soile, S., Badekas, J. Detailed 3d Courses. representations of archaeological sites. IAPRS, Vol. XXXIII,

WG V/5, Amsterdam 2000.

J.Finat Codés, M.A.de la Iglesia Santamaría, J.Martínez Rubio, J.J.Fernández Martín, J.I.San José Alonso. The Roman Theatre of Clvnia: hybrid strategies for applying virtual reality on laser scanning 3d files. ISPRS Commission V, WG V/2. Mestre (Venice), 2005.

References from Books: ASPRS, 2004: Manual of Photogrammetry. Fith Edition.

Hartley, r. and Zisserman, A.: Multiple View Geometry in Computer Vision, Cambridge University Press, 2000.

Kraus, K. 1993. Advanced Methods and Applications. Vol.2. Fundamentals and Estándar Processes. Vol.1. Institute for Photogrammetry Vienna University of Technology. Ferd. Dummler Verlag. Bonn.

References from Other Literature: Berladin J.A., Picard M., El-Hakim J., Godin G., Valzano G., Bandiera A., and Latouche D. Virtualizing a Byzantine crypt by combining high-resolution texturas with laser scanner 3D data. In Proc. VMMS 2002, pages 3-14, 2002.

El-Hakim, S.F., Beraldin, J.-A., Gonzo, L., Whiting, E., Jemtrud, M., Valzano, V. "A Hierarchical 3D Reconstruction Approach for Documenting Complex Heritage Sites. The XX CIPA International Symposium. Torino, Italie. Du 26 septembre au 1 octobre 2005. NRC 48229.

Gemenetzis, D., H. Georgiadis and P. Patias, 2001, Virtuality and Documentation: Recreating the Byzantine Heritage, Proc.